Accurate Pseudo-Constructive Wirelength and Congestion Estimation

1. Supported by MARCO GSRC Accurate Pseudo-Constructive Wirelength and Congestion Estimation Andrew B. Kahng, UCSD CSE and ECE Depts., La Jolla Xu Xu, UCSD CSE Dept., La Jolla

2. Outline Wirelength and Congestion Estimation Previous Work and Our Contribution New Wire Density Model Estimate Detoured Nets Experimental Confirmation Conclusion and Future Work

3. Wirelength and Congestion Estimation Wirelength and Congestion Estimation Problem Given:A placed VLSI standard-cell design Estimate:Total wirelength and congestion Wire density Dt(x,y) = probability that line segment  BB(t) will be used in the routed path (n, m) BB(t) (0, 0) BB(t) = Bounding box of net t

4. Outline Wirelength and Congestion Estimation  Previous Work and Our Contribution New Wire Density Model Estimate Detoured Nets Experimental Confirmation Conclusion and Future Work

5. Previous Works and Our Contribution • Lou et al. (ISPD-2001) assume that every path has the same probability of occurrence Typical Path We assume that paths with smaller number of bends have larger probability of occurrence Typical Path

6. Previous Works and Our Contribution • Previous probabilistic methods ignore detouring We take detouringinto account so that the predicted congestion map can better fit actual routing results Detoured Path

7. Previous Works and Our Contribution • Previous congestion map constructions are “one-shot”   Routing probability distributions are the same. We give an iterative congestion map construction by considering the interaction between nets.

8. Outline Wirelength and Congestion Estimation Previous Work and Our Contribution  New Wire Density Model Estimate Detoured Nets Experimental Confirmation Conclusion and Future Work

9. New Wire Density Model Goal: Set up a more practical model for congestion estimation • Instead of assuming that all paths have the same probability of occurrence We assume that only paths with the same number of bends have the same probability of occurrence • Only consider horizontal line segments • Blockage effects and detouring not included (will be handled later) pb= Probability of b-bend paths η= 0.6

10. New Wire Density Model b=1 b=2 b=3 b=4 w = Multi-bend factor

11. Blockage Effect Model Every path must pass exactly one of line segments in S. Blockage t’ x, y S

12. Test New model 20 20

13. Outline Wirelength and Congestion Estimation Previous Work and Our Contribution New Wire Density Model  Estimate Detoured Nets Experimental Confirmation Conclusion and Future Work

14. Two Questions to Answer • Which nets will detour? • How to predict detoured wirelength?

15. Relationship Between Congestion and Detouring Detoured nets are around congested regions. 200 Layout Only detoured nets are counted 200

16. Congestion_Factor F(t) = Half-perimeter of BB(t) C(x, y) = Congestion of (x, y) • Good indictor for which nets • will detour • Hard to estimate the • detoured wirelength, since • detoured wirelength also • depends on the region • outside BB(t).

17. Iterative Estimation Method • If Congestion_Factor(t) > a Expand the net bounding box in the least congested direction in order to reduce Congestion_Factor(t) Recalculate wire density function and Congestion_Factor(t) for detoured nets Keep expanding until Congestion_Factor(t) < a for all nets • Need a model for detoured nets expand

18. Detoured Nets Model • Main idea: mapping the detoured paths in an n x m grid to paths without detouring in an expanded (n+2l) x m grid. (n+l, y0) l x=n+l must be passed. (x, y)  (x, y) if y<y0  (n+2l-x, y) otherwise l = Detoured wirelength

19. Detoured Nets Model b=2 b=3 b=4 pd,b= Probability of b-bend detoured paths

20. Congestion_Factor Based Algorithm Input :Placed netlist with fixed pin location Output : Total wirelength and congestion map For each net in the layout For each line segment calculate wire density and update its congestion For each net t in the layout calculate Congestion_Factor(t) While (the maximum of Congestion_Factor > a) expand BB(t) by 1 grid in the least congested direction recalculate the Congestion_Factor(t)

21. Outline Wirelength and Congestion Estimation Previous Work and Our Contribution New Wire Density Model Estimate Detoured Nets  Experimental Confirmation Conclusion and Future Work

22. Experimental Setup • Five industry designs obtained as LEF/DEF files • Divide the layout into 40000 equal regions • Output includes total wirelength and congestion • Output results are compared with actual routing results of a commercial detailed router, Cadence WarpRoute

23. Test case RSMT_WL Actual_WL Congestion_Factor Based Est. WL improve CPU A 4.249 4.389 4.381 94% 15.69 B 3.214 3.317 3.307 78% 7.64 C 5.345 5.507 5.517 97% 3.37 D 2.190 2.372 2.368 93% 23.51 E 6.673 6.821 6.808 89% 7.91 Comparison of Total Wirelength

24. Testcase A B C D E Lou’s Model m 0.903 0.878 0.911 0.923 1.235 s 1.227 1.352 1.124 0.946 1.672 CPU 15.24 7.56 4.67 22.65 8.13 New Model m 0.951 1.037 0.962 0.981 1.102 s 1.134 0.621 0.721 0.545 1.024 CPU 6.89 3.12 1.36 9.27 3.28 New Model +Con_Fac m 0.963 1.031 0.991 0.987 1.057 s 0.689 0.566 0.707 0.323 0.813 CPU 15.69 7.64 3.37 23.51 7.91 Comparison of Estimation Quality Ideal values: m=1 and s =0 Metrics proposed in Kannan et al. DAC-2002

25. Comparison of Congestion Maps Actual Estimated 200 200 Layout Layout 200 200

26. Outline Wirelength and Congestion Estimation Previous Work and Our Contribution New Wire Density Model Estimate Detoured Nets Experimental Confirmation  Conclusion and Future Work

27. Conclusions / Future Work • New, accurate wirelength and congestion estimation methods • Improve the wirelength estimation accuracy by 90% on average with respect to the traditional RSMT wirelength estimate • More accurate congestion maps than current estimation methods • Include new estimator into a placer